Multilingual Imputation Using Transfer Learning for Estimating Emotion from Speech

Abstract

Recently, vocal communication robots attract people thanks to development of AI and robot engineering. The technology of estimating emotion from speech is important to realize a smooth dialog between human and robots. This technology needs a large number of emotional speech data, but it is difficult to collect such data a lot. We investigated the effectiveness of multilingual imputation by transfer learning using 1D convolutional bidirectional LSTM. In this paper, we report the result. The result is suggested that increasing the number of languages of emotional speech learned may exceed the performance of the model learned insufficient emotional speech in single language.